Skip to content

GEMV benchmark using fp32 #17

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
kyolebu wants to merge 1 commit into
base: main
Choose a base branch
from
Open

GEMV benchmark using fp32 #17

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions benchmarks/benchmark_gemv.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,9 +30,9 @@
)
)
def benchmark_gemv(M, K, provider):
A = torch.randn((M, K), device=DEVICE, dtype=torch.float16)
Copy link
Collaborator

@cmagina cmagina Aug 4, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Its probably worth keeping the float16 tests as well. Instead of overwriting the test, clone it and have a benchmark for fp16 and fp32.

Copy link
Contributor Author

@kyolebu kyolebu Aug 4, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Will do 👍

x = torch.randn(K, device=DEVICE, dtype=torch.float16)
y = torch.randn(M, device=DEVICE, dtype=torch.float16)
A = torch.randn((M, K), device=DEVICE, dtype=torch.float32)
x = torch.randn(K, device=DEVICE, dtype=torch.float32)
y = torch.randn(M, device=DEVICE, dtype=torch.float32)
alpha = 0.42
beta = 10
quantiles = [0.5, 0.2, 0.8]
Expand Down
51 changes: 41 additions & 10 deletions tritonblas/level2/gemv.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,38 +7,70 @@ def get_autotune_config():
return [
triton.Config(
{
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_K": 64,
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_K": 128,
},
),
triton.Config(
{
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_K": 32,
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_K": 128,
},
),
triton.Config(
{
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_K": 32,
"BLOCK_SIZE_K": 128,
},
),
triton.Config(
{
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_K": 64,
},
),
triton.Config(
{
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_K": 32,
},
),

triton.Config(
{
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_K": 128,
},
),
triton.Config(
{
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_K": 256,
},
),
triton.Config(
{
"BLOCK_SIZE_M": 8,
"BLOCK_SIZE_K": 128,
},
),

triton.Config(
{
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_K": 64,
"BLOCK_SIZE_K": 256,
},
),
triton.Config(
{
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_K": 512,
},
),
triton.Config(
{
"BLOCK_SIZE_M": 8,
"BLOCK_SIZE_K": 256,
},
),
]
Expand Down Expand Up @@ -90,8 +122,7 @@ def gemv_kernel(
tiled_x = tl.load(x_block, mask=k_mask, other=0.0)

# Partial dot product
partial_result = tl.sum(tiled_a * tiled_x[None, :], axis=1)
accumulator += partial_result
accumulator += tl.sum(tiled_a * tiled_x[None, :], axis=1)

a_block += BLOCK_SIZE_K * stride_ak
x_block += BLOCK_SIZE_K * stride_x
Expand Down